Depending upon your philosophical viewpoint, either tmpfs always does
direct IO, or it cannot ever do direct IO; but whichever, if tmpfs is to
stand in for a more sophisticated filesystem, it can be helpful for tmpfs
to support O_DIRECT. So, give tmpfs a shmem_direct_IO() method, of the
simplest kind: by just returning 0 done, it leaves all the work to the
buffered fallback (and everything else just happens to work out okay -
in particular, its dirty pages don't get lost to invalidation).
xfstests auto generic which were not run on tmpfs before but now pass:
036 091 113 125 130 133 135 198 207 208 209 210 211 212 214 226 239 263
323 355 391 406 412 422 427 446 451 465 551 586 591 609 615 647 708 729
with no new failures.
LTP dio tests which were not run on tmpfs before but now pass:
dio01 through dio30, except for dio04 and dio10, which fail because
tmpfs dio read and write allow odd count: tmpfs could be made stricter,
but would that be an improvement?
Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
---
mm/shmem.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/mm/shmem.c b/mm/shmem.c
index 7420b510a9f3..4d5599e566df 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -2720,6 +2720,16 @@ shmem_write_end(struct file *file, struct address_space *mapping,
return copied;
}
+static ssize_t shmem_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
+{
+ /*
+ * Just leave all the work to the buffered fallback.
+ * Some LTP tests may expect us to enforce alignment restrictions,
+ * but the fallback works just fine with any alignment, so allow it.
+ */
+ return 0;
+}
+
static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
{
struct file *file = iocb->ki_filp;
@@ -4421,6 +4431,7 @@ const struct address_space_operations shmem_aops = {
#ifdef CONFIG_TMPFS
.write_begin = shmem_write_begin,
.write_end = shmem_write_end,
+ .direct_IO = shmem_direct_IO,
#endif
#ifdef CONFIG_MIGRATION
.migrate_folio = migrate_folio,
--
2.35.3
